Windows sandbox failure wasted user time and Codex usage while blocking all local work

Windows Codex desktop sandbox failure wasted user time and Codex usage while blocking all local work

User impact

This failure wasted a substantial amount of the user's time and Codex usage/credits. The user was trying to perform a normal repository workflow: review a report and save a Markdown file. Because the local execution layer failed, the task required repeated retries, permission changes, task refreshes, and repeated explanations, while no local file operation could complete.

The user explicitly reports that this failure consumed a significant amount of time and quota without producing the requested local artifact.

Environment

• OS: Windows 11 x64
• Surface: Codex desktop app
• Date: 2026-09-20 (Asia/Shanghai)
• Effective task sandbox shown to the agent: workspace-write
• Writable root shown to the agent: C:\Infinite-Canvas
• Approval policy shown to the agent: never

Reproduction

1. Open a local project in Codex desktop.
2. Select the maximum/full-access permission profile.
3. Ask Codex to run a trivial local PowerShell command.
4. Ask Codex to apply a file patch inside the writable workspace.
5. Repeat after refreshing/reopening the task.

Actual result

Every local command fails before the requested process starts:

exec_command failed: CreateProcess {
  message: "Rejected(\"Failed to create unified exec process: helper_unknown_error: setup refresh had errors\")"
}

File patches fail at the same sandbox setup layer. The failure is not limited to one repository or one file, and it prevents both read-only inspection and writes.

The full-access selection does not recover the active task. The task continues to expose workspace-write and the Windows sandbox helper still fails during setup refresh.

The in-app browser/Node runtime also exits with:

windows sandbox failed: helper_unknown_error: setup refresh had errors

Expected result

Full access should either bypass the Windows filesystem sandbox or initialize it successfully. A sandbox initialization failure should not silently leave the agent unable to perform local work while consuming user turns and usage quota. The error should expose the underlying Windows failure and provide a recovery path.

Impact

The agent could still produce chat responses and browse the web, but could not:

• inspect local files reliably;
• edit or create a Markdown report;
• run tests;
• complete a repository workflow.

This makes the coding agent effectively unusable for local work and causes avoidable user time and quota loss.

Related reports

This appears related to:

• https://github.com/openai/codex/issues/44696
• https://github.com/openai/codex/issues/33964
• https://github.com/openai/codex/issues/24416

Please investigate the Windows sandbox setup-refresh failure, the failure to propagate the full-access permission profile to the active runtime, and whether failed tool attempts are incorrectly consuming user usage/credits. Please also provide a recovery path that does not require repeated retries or task recreation.

The issue names no source files or tests. Start by reproducing the failing PowerShell command and tracing the Windows sandbox setup-refresh path that emits helper_unknown_error; done means local inspection, commands, and file patches work under the selected permission profile, with an actionable recovery error when setup fails.